On 08/18/2014 05:51 AM, Christoffer Dall wrote: > On Wed, Aug 13, 2014 at 06:20:19PM -0700, Mario Smarduch wrote: >> On 08/13/2014 12:30 AM, Christoffer Dall wrote: >>> On Tue, Aug 12, 2014 at 06:27:11PM -0700, Mario Smarduch wrote: >>>> On 08/12/2014 02:50 AM, Christoffer Dall wrote: >>>>> On Mon, Aug 11, 2014 at 06:25:05PM -0700, Mario Smarduch wrote: >>>>>> On 08/11/2014 12:13 PM, Christoffer Dall wrote: >>>>>>> On Thu, Jul 24, 2014 at 05:56:08PM -0700, Mario Smarduch wrote: >>> >>> [...] >>> >>>>>>>> @@ -1151,7 +1170,7 @@ static void kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data) >> >> Ok I see, I was thinking you thought it was breaking something. >> Yeah I'll add the comment, in reality this is another use case >> where a PMD may need to be converted to page table so it makes sense >> to contrast use cases. >> > > the hidden knowledge is that MMU notifiers will ensure a huge PMD gets > unmapped before trying to change the physical backing of the underlying > PTEs, so it's a gigantic kernel bug if this gets called on something > mapped with a huge PMD. > That's a good way of putting it luckily I was aware previously looking at some other features. > >>> >>>> >>>> Should I add comments on flag use in other places as well? >>>> >>> >>> It's always a judgement call. I didn't find it necessarry to put a >>> comment elsewhere because I think it's pretty obivous that we would >>> never care about logging writes to device regions. >>> >>> However, this made me think, are we making sure that we are not marking >>> device mappings as read-only in the wp_range functions? I think it's >> >> KVM_SET_USER_MEMORY_REGION ioctl doesn't check type of region being >> installed/operated on (KVM_MEM_LOG_DIRTY_PAGES), in case of QEMU >> these regions wind up in KVMState->KVMSlot[], when >> memory_region_add_subregion() is called KVM listener installs it. >> For migration and dirty page logging QEMU walks the KVMSlot[] array. >> >> For QEMU VFIO (PCI) mmap()ing BAR of type IORESOURCE_MEM, >> causes the memory region to be added to KVMState->KVMSlot[]. >> In that case it's possible to walk KVMState->KVMSlot[] issue >> the ioctl and come across a device mapping with normal memory and >> WP it's s2ptes (VFIO sets unmigrateble state though). >> >> But I'm not sure what's there to stop someone calling the ioctl and >> install a region with device memory type. Most likely though if you >> installed that kind of region migration would be disabled. >> >> But just for logging use not checking memory type could be an issue. >> > I forgot that the current write-protect'ing is limited to the memory > region boundaries, so everything should be fine. I looked through this way back, but it was worth to revisit. > > If user-space write-protects device memory regions, the worst > consequence is that it breaks the guest, but that's its own > responsibility, so I don't think you need to change anything. Thanks for the detailed review. I'll go off now, rebase and make the needed changes > > -Christoffer > -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
